WO2017021644A1

WO2017021644A1 - Processing of geometric data with isotopic approximation within a tolerance volume

Info

Publication number: WO2017021644A1
Application number: PCT/FR2016/051997
Authority: WO
Inventors: Pierre Alliez; David COHEN-STEINER; Manish MANDAD
Original assignee: Inria Institut National De Recherche En Informatique Et En Automatique
Priority date: 2015-08-01
Filing date: 2016-08-01
Publication date: 2017-02-09
Also published as: FR3039686B1; EP3329465A1; US20190019331A1; FR3039686A1; FR3039685A1

Abstract

The processing of geometric data starts with a tolerance volume Ω that is related to raw geometric starting data. A canonical triangulation is initiated (201) in the tolerance volume, on the basis of a set S of sample points taken at the limits of this tolerance volume, and points LBB located on the exterior thereof. This canonical triangulation is refined (203) until the sample points have been classified, thereby obtaining a dense mesh of the tolerance volume. This dense mesh is then simplified (207) via triangulation modification operations, while preserving the topology and classification of the sample points.

Description

Geometric data processing, with isotopic approximation in a tolerance volume.

The invention relates to computer techniques for constructing or reconstructing geometric surfaces. At the base, a surface or shape can be represented by a set of points in space. This is called a "point cloud".

To enable computer processing as a surface, a two-dimensional surface is often represented by a mesh based on triangular primitives. Similarly, a three-dimensional volume can be represented by a tetrahedral mesh. The expressions "triangular mesh" and triangulation cover all of these cases.

A representation of surfaces by "triangular mesh" or parametric surfaces can be imperfect, most often because of the limits of the sensors used to establish it and the algorithms of surface reconstruction from measurements. Faults can be eg holes, or auto-intersections (because it is rare that a surface of the real world crosses itself). The expression "soup of triangles" is used to aim for a triangulation that may be imperfect.

Realizing a faithful approximation of complex shapes with triangular meshes (also called simplicial) is a multifaceted problem, which involves geometry, topology, and their discretization (division into primitive).

This problem has attracted considerable interest because of its wide range of applications and the ever-increasing availability of geometric sensors. This results in significantly increased availability of scanned geometric models, which does not mean that their quality is increasing: while many practitioners have access to high precision acquisition systems, the recent trend is to replace these expensive tools with heterogeneous combination of consumer level acquisition devices. The measures generated by the latter are insusceptible to a direct treatment. This is even more true if you want to merge heterogeneous data from different consumer sensors. Similarly, the diversity of geometric processing tools results in an increase in the number of defects in the data: the conversion to and from different geometrical representations often has the effect of losing quality in relation to the input signal. Few algorithms provide better quality guarantees on their output than their input signal requirements.

More and more discretizations are used to capture complex geometric shapes. This makes the concern become dominant to offer strict geometric guarantees, to be robust to the appearance of artifacts and make possible the certification of algorithms.

"Geometric guarantees" means the fixing of a maximum to the errors of the geometric approximation, and the absence of auto-intersections.

In addition, there are topological guarantees, such as homotopy, homeomorphism or isotopy. These notions are defined below.

In this case, isotopy means that there is a continuous deformation that deforms one form over another, while maintaining a homeomorphism between these two forms.

Surface meshes with these warranties are required for artifact-free rendering, digital engineering, reverse engineering, manufacturing, and 3D printing.

Geometric simplification reduces the number of primitives. Topological simplification can repair holes and degenerations present in existing discretizations. These two types of simplifications can be used in combination to reconstruct clean forms from raw geometric data such as point clouds, or triangle soups.

Numerous techniques have been proposed over the years to make computer approximations of shapes. Few of them provide an upper limit for errors.

Document [Refl] (Appendix 2) proposed a method of approximation in polynomial time, with a guarantee of going down to a minimum number of vertices for a given maximum error tolerance. But this result is purely theoretical, and this document does not offer a practical algorithm for implementation. The maximum error approximation has been sought also by clustering in [Refl8], by mesh decimation [Ref9], [Refl5], [ReflO], [Ref28] or by a combination of these two techniques. [Ref24].

In general, the error metric considered is the unidirectional Hausdorff distance that goes back to the input mesh, but some have used bidirectional distance [Ref 9].

In general, these approaches are not generic enough to work on heterogeneous and imperfect input data. And they do not guarantee an exit without intersections.

One can avoid creating intersections during the decimation of meshes ([Ref26]), but this is not enough in the case of input data which themselves include auto-intersections, and these avoidances imply an early end to the simplification phase, which ultimately results in complex meshes.

In general, it is important to output a mesh that is isotopic to the input data, and is content with a small number of triangles (or vertices). In 2D, this is called the minimal registered polygon problem, and has been well studied in the convex case [Ref22]. In 3D this becomes the problem of minimal inscribed polyhedra, which has been shown to be of N-P difficulty. Although this problem is old, there is still no robust and practical solution to this scientific and technical challenge. The present invention provides a remedy.

The method proposed for this purpose leaves, as input data, a tolerance volume Ω. This tolerance volume can be determined, as part of a pre-treatment, from raw geometric raw data starting, for example on the basis of a cloud of points or a soup of triangles, or from any set of geometric data, even heterogeneous. The proposed method comprises the following steps:

A. receive a tolerance volume Ω relative to raw geometrical data of departure,

B. initiating a canonical triangulation in the tolerance volume, from a set S of sample points taken at the limits of this tolerance volume, and points situated outside of it,

C. refine this canonical triangulation, to classify the points of the sample, which provides a dense primary grid, of correct topology, for the volume of tolerance, and

D. simplify this dense mesh by triangulation modification operations, preserving the topology and the classification of the points of the sample.

Also provided is a geometric data processing device comprising:

A geometric data memory, for receiving graphic data representing a tolerance volume,

An initialization member, arranged to perform a first canonical triangulation between the edge of the tolerance volume and points located outside thereof,

A refinement organ, arranged to specify said first triangulation to provide a dense mesh of the tolerance volume, which preserves a classification condition of the points of the triangulation,

A simplification element, arranged to simplify said dense mesh, preserving the topology and the classification of its points, and a pilot which successively actuates the initialization member, the refinement member, and the simplification member, for changing graphic data from the same tolerance volume. Other features are also proposed, for the process as for the corresponding device.

The initialization step B may comprise the sub-step of sampling the limits of the tolerance volume according to a selected sampling density σ, which provides the set S of sample points.

Step B may include marking each sample point of the set S with a tag value that represents the index of the edge connected component of that point in the tolerance volume. The refinement step C then comprises the iterative generation of the canonical triangulation, by inserting one point at a time, on at least a part of the set S of points, until all the points concerned verify a condition of classification, based on the value of a function F itself depending on their index value, and on its interpolation by an interpolation function f applied to each mesh of the 3D triangulation.

Thus, it is possible to make use of a point classification operator, which receives the designation of a point to be treated, and determines an interpolation function f on the basis of the mesh surrounding this point to be processed, and of values labels relating to the vertices of this mesh, to provide an interpolated label, and to compare it with a current label of the point to be treated, which makes it possible to determine a condition of classification of the point to be treated. Step C can be completed by the construction of a piecewise linear isosurface Zset, using the points where the interpolation function f has the same selected intermediate value, in particular zero, this Zset isosurface being a closed surface. , which forms a surface approximation of the volume of tolerance.

Preferably, the canonical triangulation is a Delaunay triangulation. The raw geometric raw data can be 2D data or 3D data.

For its part, the simplification step D. can comprise the following operations:

Dl. determining whether triangulation modification operations are valid, that is, preserving the topology of the mesh and preserving the condition of classifying the set S of sample points, and

D2. perform valid triangulation modification operations in a specified order.

Preferably, the triangulation modification operations comprise edge fusions and / or half-edge fusions. The order can be determined according to an optimization criterion.

In one embodiment, step D. relates to a so-called "simple tolerance" volume that approaches the tolerance volume by a mesh of the canonical triangulation, the Zset isosurface being contained in this volume called "simplistic tolerance" ,

The steps D1 and D2 can then be performed:

First on the edge of the volume called "simplistic tolerance",

Then on the fruit of a mutual triangulation between the isosurface (Zset) and the volume called "simplistic tolerance", this mutual triangulation adding vertices, edges and meshes in the triangulation, on the isosurface,

Finally, with other edges contained in the volume called "simplistic tolerance", to cover all possible edges.

The invention is also directed to a computer program product comprising portions of program code for implementing the above method or device, when the program is run on a computer.

Other features and advantages of the invention will appear on examining the detailed description below, and the attached drawings, in which:

- Figures 1-a to 1-e schematically illustrate a first example of application of the invention;

FIG. 2 is the general flow diagram of the mechanism proposed according to the invention,

FIG. 3 illustrates an example of the beginning of application of the invention.

FIG. 4 is a flow diagram which details that of FIG. 2,

FIG. 5 is a drawing illustrating the definition of barycentric ratios,

FIG. 6 is a flow diagram which details a preferred embodiment of the refinement step 203 of FIGS. 2 and 4,

FIGS. 7 show different state of the triangulation during the refinement phase, in another example, very geometric, of application of the invention,

FIGS. 8 show different state of the triangulation during the simplification phase, on the same example, very geometric, of application of the invention; FIG. 9 are drawings illustrating a particular case of application of the invention; , FIGS. 10 are representations of elephants showing the robustness of the method according to the invention,

FIGS. 11 diagrammatically illustrate other examples of application of the invention, on the same object as FIGS.

FIG. 12 is a diagram illustrating certain advantages of the invention,

FIG. 13 is another diagram illustrating certain advantages of the invention,

FIG. 14 is another representation, called "fertility", on which the diagrams of FIGS. 12 and 13

FIGS. 15 are representations comparable to that of FIG. 3, making it possible to better understand the sub-phase of mutual triangulation, in the simplification phase,

FIG. 16 illustrates two mechanisms for melting edges and half-edges,

FIG. 17 is a flow diagram which details a preferred embodiment for the simplification step 207 of FIGS. 2 and 4; FIG. 18 is the block diagram of a device according to the invention, and

- Figure 19 is the block diagram of a classification operator used according to the invention.

The drawings and the description below contain, for the most part, elements of a certain character. They can therefore not only serve to better understand the present invention, but also contribute to its definition, if any. The nature of the invention means that certain characteristics can only be fully expressed by the drawing.

The description is followed by an Appendix 1 with definitions, and an Appendix 2 containing the coordinates of the documents referred to here. These appendices form an integral part of the description.

FIGS. 1a to 1e will firstly be examined for a very general description of the invention, with reference to FIG. 2, which illustrates the major operating steps of the proposed method, and FIG. 4, which details the FIG. 2.

Figure 1-a illustrates the starting point, namely a tolerance volume Ω relative to a mechanical part measured by a laser scanner, with an input tolerance δ of 0.6%. Figure 1-b illustrates what is obtained after the refinement step (203, Figure 2). The result is a piecewise linear function Z, corresponding to a surface mesh of the tolerance volume Ω of FIG. 1. The number of vertices is 20.4k (20,400). Simplifications are then carried out in accordance with step 207 (FIG. 2).

FIG. 1-c illustrates the same mesh as FIG. 1b, but after it has undergone a first simplification (simplification of the edge oT of the "simplistic" tolerance Γ, step 2072, FIG. 4), which makes it go down to 5 , 3k vertices.

FIG. 1-d illustrates the same mesh, after mutual triangulation and simplification of Z (step 2074, FIG. 4), which makes it go down to 1.01 k vertices.

Finally, FIG. 1-e illustrates the mesh at the end of simplification (step 2076, FIG. 4), which is further reduced to 752 vertices.

It is observed, in the eye, that, even simplified, the mesh of Figure 1-e represents the object defined by the previous figures. Reference is now made to FIG. 3, which uses a symbolic curvilinear tolerance volume, simpler than that of FIGS.

It has been seen that FIG. 2 illustrates the major operating steps of the proposed method. Figure 4 illustrates some process steps in more detail.

In FIG. 3, the tolerance volume Ω is defined by a spiral curvilinear shape, whose 0Ω boundary area has two borders όΩι and 0Ω ₂ , in a sectional view in plan.

Step 201 of Figures 2 and 4 is an initialization. As indicated in 2011 (Figure 4), the initialization first includes a sampling of the edge (or the border) of the tolerance volume. A set S of sample points 5 is obtained, stored in step 2013 (FIG. 4). This sampling is in principle "σ-dense". This means that balls of radius σ centered on all points s of S completely cover the 0Ω boundary zone of the tolerance volume Ω. In practice, σ is a fraction of the thickness δ of Ω.

In the simplest case, that of Figure 3, the 0Ω boundary area has only two connected components of edge surfaces, namely όΩι (external boundary) and 0Ω ₂ (internal boundary). In this case, it is easy to distinguish the inside and the outside from one end to the other of the 0Ω border zone. This can be done formally, by assigning each sample s of the set S a tag value (or label) F (s) which is, for example:

F (s) = 1 if the point s is on όΩι, and

F (s) = -1 if the point s is on ÔQ ₂ . This value F (s) can be seen as a representation of the index (or "index") of the connected edge component of the considered point s, in the tolerance volume.

Another function 'f is constructed: For a point x, present in a mesh M (triangular in 2D, tetrahedral in 3D), one computes a function of linear interpolation f (x, M), defined on the basis of the values F (s) the three or four vertices of the mesh M (triangle or tetrahedron), and the barycentric coordinates of the point x considered, with respect to the vertices of the mesh. More precisely, f (x, M) is a convex linear combination of the values of F at the vertices vi of the mesh M:

where for all i, eu is positive or zero, and the sum of eu is 1. The values of eu are defined for example as shown in Figure 5:

For a point X located in a triangular mesh ABC, one of the eu is the value a for X with respect to the vertex A of the triangle, denoted (A (X), which is the ratio of the surface of the triangle XBC to the surface of the triangle ABC.

We will thus have: f (x, M) = a _A (X) F (A) + a _B (X) F (B) + a _c (X) F (C)

It is the same thing with a tetrahedral mesh, which has one more vertex, and where the surfaces become volumes.

For a point X in a tetrahedral lattice ABCD, the value a for X with respect to the vertex A of the tetrahedron, denoted by (A (X), is the ratio of the volume of the tetrahedron XBCD to the volume of the tetrahedron ABCD, and the formula becomes: f (x, M) = a _A (X) F (A) + a _B (X) F (B) + a _c (X) F (C) + a _D (X) F (D)

In 3D, in the case of tetrahedra, the interpolated function f (x, M) is constrained to be a function that satisfies the so-called Lipschitz condition (having a limited derivative). The skilled person knows how to implement this additional condition, which is needed during refinement. Examples will be given later, with reference to FIGS.

Further information on these barycentric calculations is available in [Ref25].

Beside this, the set of points p where f (p) = 0 corresponds to a polygonal line (a 3D surface) called ZeroSet. This ZeroSet polygonal line, also abbreviated as Z, is a linear part-line, which is placed within the tolerance volume.

We also use an error quantity ε, which can be defined as follows: ε (s) = I f (s, M) - F (s) I, where the notation | x | aims the function absolute value of x. As part of the set S, the point s has a value F (s). In the mesh M of the mesh which contains this point s, this one receives an interpolated value f (s, M).

We consider a parameter a, with (0 <a <1). The default value of a can be set to 0.2. We will then say: if ε (s) <= (1 - a), that the point s is well classified with the margin a, - Otherwise, that the point s is misclassified.

A value ε close to 1 indicates that the mesh must be refined.

This refining is done by adding to the set Sp (set of points which are the subject of the triangulation) a point x of the set S which is not yet part of Sp. This point x inserted in the mesh has a value-label F (x). We then redo the classification calculation, and the error ε, for all the non-inserted points of the set S (that is to say for all the points of the set S that are not part of the 'set Sp).

In theory at least, it would be possible to proceed in an arbitrary order. It is currently preferred to start with the points that give the greatest error, and then to iterate. This is what will be described later, with reference to the mechanism illustrated in FIG.

It is observed that this mechanism converges whatever the mode of selection of the added points: in the worst case one would add all the points of the sample, and they are then all well classified. As part of a Delaunay refinement, the points that are added to the current set of samples to be triangulated (Sp) are called "Steiner points." (Not to to confuse with the Fermat point of a triangle, which some also call "Steiner's point").

When this is complete, and all the points are correctly classified, that is to say that the error is everywhere <= (1 - a), we have a correct, but complicated, primary mesh.

We will now begin the description based on Figures 7, which are very geometric, which facilitates understanding. These figures illustrate the two-dimensional treatment, which the skilled person can transpose in three dimensions. Only the 2D illustration is possible; indeed, 3D figures would be illegible because everything would be superimposed on the screen.

Figure 7.0 shows a tolerance volume as is.

Figure 7.1 shows the initialization phase, in which the points S of the sampling S of the edges of the border zone of the tolerance volume appear. On the outside, for the label F (s) = 1, these points are hollow circles. Inside, for the label F (s) = -1, they are solid circles.

Figure 7.2 shows a so-called extended bounding box envelope LBB. This expanded bounding box LBB (or "Loose Bounding Box") can be constructed by calculating the strict bounding box of the tolerance volume, using the minimum and maximum coordinates in x, y, z points of the tolerance volume. . This box LBB is enlarged by homothety by a factor 1.5, for example.

The expanded bounding box LBB includes the points LBB1 to LBB4 shown in Figure 7.2. In 2D, the bounding box is formed by only 4 points, against 8 points in 3D. Other means can be used to define external points whose embedding property is guaranteed. We then consider another set Sp of points s _p . Initially empty, this set Sp first receives all the points of the extended bounding box LBB.

In step 2015 of FIG. 4, one starts by making a triangulation T from these first points of the LBB placed in Sp, but which are not points of S.

The line segments that form the boundaries of the stitches are shown in long dashed lines in FIGS.

The points of the LBB are given special label values, which depend on their distance from the edge of the tolerance volume. These special tag values can be greater than 1. For example: the label-value 1 is assigned to the outermost edge of the tolerance volume, and the label 'd + to the vertices of the LBB, where d is the normalized distance between a vertex of the LBB and the point of the outer edge the tolerance volume that is closest to the top of the LBB. Normalization is a function of the tolerance volume thickness, which is determined at value 2 (the difference between -1 and +1).

We then make the classification calculation, and the error ε, for all non-inserted points of the set S, to which are added the points of the extended bounding box which are in the set S _p . This calculation suffers an exception in the case where a point of the sample is located in a mesh where the function is homogeneous, that is to say with labels identical to the vertices of the mesh. In this case if the function has the same sign as the label of the point of the sample then it is well classified, and badly classified otherwise, without using the error ε. As already indicated, all the points s of the set S have been assigned a label value F (s). The interpolation function f is constructed, as before, for each cell (2D triangle, 3D tetrahedron): For a point x of the set S _p , present in a tetrahedral cell, a linear interpolation function is computed f (x, M), defined on the basis of the label values F (s) at the vertices of the tetrahedron mesh M, and the barycentric coordinates of the point x considered, with respect to the vertices of the mesh. And each time the error quantity ε is calculated, which indicates whether the point x is or is not well classified.

In Figure 7.2, a well-classified point is represented by a square. A badly classified point is represented by a cross. It is observed that the interpolated function on the two triangles LLB1-LBB2-LBB3, as well as LLB1-LBB3-LBB4 is here homogeneous and positive in all points, since the vertices of the LBB all have a label 'd + which is clearly positive.

We are therefore within the framework of the above exception.

The hollow circles of Figure 7.1 have a + 1 positive label. They are therefore of the same sign as the interpolated function on the two triangles, and therefore well classified. Therefore, these hollow circles in Figure 7.1 become squares in Figure 7.2.

For their part, the full circles of Figure 7.1 have a negative label - 1. They are therefore of opposite sign to the interpolated function on the two triangles, and therefore poorly classified. As a result, these solid circles in Figure 7.1 become crosses in Figure 7.2. Then begins the refinement itself (step 203 of Figure 2, and frame 203 of Figure 4).

In step 2031 of FIG. 4, a first point of Steiner SP1 is added in the triangulation. This point SP1 is added to the set S _p of the points to be triangulated. This point SP1 may be the one whose error ε is maximum, as indicated here elsewhere.

In figure 7.3, this first point of Steiner SPl happens to be close to the diagonal L1 of the envelope LBB. The label of SPl is F (SP1) = -1.

We then redo the classification calculation, and the error ε, for all the points of the set S, and for the added point SP1 (step 2032 of FIG. 4).

The two extreme points LBB1 and LBB3 of the diagonal Ll are part of the LBB. This determines a version "SPl" of the function f, calculated as indicated above, in the 4 triangular meshes LBB1-LBB2-SP1, LBB2-LBB3-SP1, LBB1-LBB4-SP1, LBB3-LBB4-SP1. We finally obtain f (SPl) = -1, which classifies well the added point SPl. Around the point SP1 is then defined an isosurface IS1, where f = 0. This isosurface is illustrated by a black broken line in bold lines, almost square, in figure 7.3.

Consider the triangle LBB1-LBB2-SP1. The interpolated function is -1 in SPl, and a positive value greater than 1 in LBB1 and LBB2, say 5. Along the edge SP1-LBB1 the function therefore varies linearly from -1 to 5, and so we find the Zero level of the interpolated function f near SPl, above.

Now consider the triangle LBB3-LBB4-SP1. By symmetry, we find another zero level of the interpolated function f near SPl, below, this time.

The two vertical segments of IS1 are explained in the same way, considering the triangles LBB2-LBB3-SP1 and LBB1-LBB4-SP1. The 2D analysis made above can be extrapolated to 3D in the following way: In 3D, the interpolated function f is always linear per piece, but evolves in space so there is a depth in addition; and the isosurface zeroset (or Zset) is no longer a segment in a triangle, but a planar polygon in a tetrahedron, namely a triangle or a quadrilateral. We observe that, on the horizontal line of points contained in the center of ISI, the points are now well classified, and the crosses have become squares, with F (s) = 1. On the contrary, the points situated outside of They have remained crosses, with F (s) = -1, therefore misclassified. It is observed that the lower part of the contour IS1 is halfway between the horizontal line of points contained in IS1, and the line of cross below.

The other elements of the IS1 contour depend on the "SP1" version of the function f, as explained above. The isovalue IS1 represents the zero level of the interpolated function, which is therefore positive on one side of IS1, and negative on the other side of IS1. The points of the sample are well classified when they are located on the side of ISl, which corresponds to the sign of their label or value-label.

Other useful information to build a Delaunay triangulation can be found in [Ref23]. This information is to be considered as incorporated in the present description.

It should be noted that, for points in general position (in 2D, no more than 3 co-circular points, and in 3D no more than 4 co-spherical points), there is only one Delaunay triangulation, which is therefore canonical . A property is in 2D that the circles circumscribed to the triangles contain no other point; similarly, in 3D, the spheres circumscribed in the network tetrahedra (constituted by points of the subset Sp of S) are empty, and contain no point other than the vertices of the tetrahedron.

The Delaunay triangulation is completed when all the points of the set S are correctly classified (step 2033 of FIG. 4).

If this is not the case, steps 2031 to 2033 of FIG. 4 are repeated, inserting a new Steiner point.

This is done with point SP2 of Figure 7.4, which is added to the set S _p of the points to be triangulated. We then redo the classification calculation, and the error ε, for all the points of the set S, and for the points SP1 and SP2 added (step 2032 of FIG. 4). This results in a so-called "SP2" version of the function f.

The points where this "SP2" version of the function f is zero define a new isosurface IS2, which includes the points SP1 and SP2. On the CC20 and CC22 segments of the tolerance volume contour, the points remain poorly classified (crosses). On the other hand, in the vicinity of SP2, the segments CC21 and CC23 contained in IS2 now include well-classified points.

Similarly, the quadrant CC24 becomes well classified, unlike segments aligned on him, but outside of IS2. Conversely, the CC25 and CC26 square segments become misclassified, unlike the segments aligned on them, but outside IS2.

In the test step 2033 of FIG. 4, certain points of the set S remain poorly classified. Thus, steps 2031 to 2033 of FIG. 4 are repeated by inserting a new Steiner point SP3, as illustrated in FIG. 7.5.

We then redo the classification calculation, and the error ε, for all the points of the set S, and for the points SP1, SP2 and SP3 added (step 2032 of FIG. 4). This results in an "SP3" version of the f function. The points where this "SP3" version of the function f is zero define a new isosurface IS3, which includes the points SP1, SP2 and SP3.

As before, we observe that the points that are inside IS3 do not have the same classification as the points located on the same line as them, but outside of IS3. In the test step 2033 of FIG. 4, certain points of the set S remain poorly classified.

Steps 2031 to 2033 of FIG. 4 will be repeated several times, each time inserting a new Steiner point. This gives the figure 7.6, where Steiner's points are big black circles, if their value-label is negative, or big inner circles white, if their value-label is positive. The isosurface obtained approximates the general shape of the volume of tolerance, but there are still badly classified points.

Steps 2031 to 2033 of FIG. 4 will be repeated several times, each time inserting a new Steiner point. This leads to Figure 7.7, where the isosurface is entirely contained in the Tolerance Volume, and there are no more crosses (misclassified points).

At the test step 2033 of Figure 4, the end of the refinement has now been reached.

So we go to phase 207 of simplification (or decimation), Figures 2 and 4.

Step 203 of Figures 2 and 4, which is called "refinement", has been described. One of the goals of this refinement is to arrive at a mesh without surface intersection, which corresponds to the topology of the volume of tolerance. As indicated by the word "refinement", one makes evolve the mesh of "coarse-to-end", with the points successively inserted. The mesh itself can be done using a Delaunay triangulation, whose data structures do not accept intersecting surfaces. Other techniques can be used to construct canonical triangulations that do not intersect. The Applicant currently prefers a triangulation using the Delaunay property, with the proof of termination and the topology guarantee that it allows.

As indicated, the selection of inserted Steiner points can be done according to the error magnitude ε. This can be done on the basis of the mechanism of Figure 6, which will now be described in 3D version. In step 601, for each tetrahedron of the triangulation T, a list of those sample points s of the set S that are contained in this tetrahedron is maintained. Each point is accompanied by its label, and its error ε. The points of the set S _p that are already included in the triangulation are not in this list. In step 603, a dynamic priority queue is maintained for each tetrahedron of the triangulation T, containing the maximum error points in the tetrahedrons, sorted by decreasing error.

In step 605, which corresponds to step 2031 of FIG. 4, the maximum tail end error point is added to triangulation T, which can then be deleted from the tail.

In step 607, which corresponds to step 2032 of FIG. 4, the error calculations are updated. A test 609, which corresponds to step 2033 of FIG. 4, examines whether the topology is correct, i.e. if the set S is completely classified. If this is not the case, we return to 601. Otherwise, we can proceed to the next phase of simplification of the mesh.

At the end of refinement, the mesh representation of FIGS. 7.6 and 7.7 is correct, in particular in that it does not include intersections, and that it is isotopic. On the other hand, it is too dense, and contains an excessive number of meshes (or vertices). We will simplify it, as we will see later. We will first consider Figure 7.8.

At the end of the triangulation mesh that constitutes refinement, we have a set of tetrahedra. The "simple tolerance", denoted by Γ, is the union of the tetrahedra which constitutes a meshed (triangulated) approximation of the tolerance volume Ω. These tetrahedra have vertices with different labels. The edge of this "simple tolerance" is noted dT (in Figure 7.8). In 3D, this edge is composed of triangles which are the free faces of the mesh tetrahedra. It is a triangulated representation of the boundary of the tolerance domain. The mixed line in Figure 7.8 illustrates those edges of this triangulated representation that are in the plane of the figure.

From there, the simplification of the mesh representation of FIGS. 7.7 or 7.8 can be done essentially by using a process known as "edge collapse".

The nature of this process will now be described with reference to Figs. 16, which are in 2D.

Figure 16.0 shows the ends A and B of an edge that can be fused (we will see the conditions later).

The general merge operator deletes the edge that joins the vertices, at least one of these two vertices, and the faces or tetrahedra adjacent to the removed edge (Figure 16.2). The position of the vertex P resulting from the fusion (target vertex) is a degree of freedom (continuous) of this operator, under condition to preserve the topology and to classify all the points well.

In the case where these two constraints (topology and classification) can be satisfied, there are one or more regions of the space (region of validity) where the target vertex P can be placed. It can be arbitrarily placed in this region, but it is better to choose an optimum point.

For this, one-point error calculation can be used. The error is, for example, the sum of the squares of the distances from this point to the support planes of the facets of the isosurface Z, these facets being selected in a local neighborhood at the considered edge. This local neighborhood may include facets of the isosurface that are located at a distance 2 from the edge in the triangulation graph (so-called "2-ring"). We look for the optimum point that minimizes this error.

This error calculation serves as an optimization criterion. Indeed, as preferred point P, we can then take the point located in the region of validity, and which is as close as possible to the optimum point that minimizes said error.

It should be noted that this mode of error calculation and determination of the preferred point P is particularly relevant for planar surfaces per part; However, we can imagine other solutions depending on the nature of the input surfaces. For example, for curved surfaces in parts, we can imagine choosing no longer planes but parametric surfaces of the quadric or cubic type, and to take as error the sum of the squares of the distances to these surfaces.

In practice, to increase the speed of the processing, it is interesting to use also, as a variant, a purely combinatorial operator that is called a "half-edge fusion operator": in this less general variant, the operator is forced to merge one vertex of the edge on the other. In FIG. 16.1, the vertex A is thus fused to the sum B. The optimization criterion is defined by the distance between the target point (here the point B) and the optimum point Q. calculated as above. .

This half-edge merge operator is less powerful than the edge merge operator that admits to placing the vertex resulting from the merge at a position that generates the P in space, but it is much faster to simulate, since we directly know the target point B, without having to look for the best point P.

The half-edge fusion operator is called combinatorial in the sense that there exists a finite number of degrees of freedom: 2 E possible fusions, where E represents the number of edges of the triangulation. Then, when no half-edge fusion is possible, one uses the more general operator of fusion of edges.

The difference between the two phases should be clearly understood: refinement and simplification.

During refinement, sample points are inserted in a canonical triangulation (whose connectivity is determined by the positions of the points). So, in this phase of refinement, the degrees of freedom are only in terms of choice of points.

During the simplification, we will see that the merge operators act by edges or half-edges, so a priority queue sorts the operators that are valid (which preserve the topology and classification of the points), following a score that depends on a optimization criterion.

The ultimate goal is to simplify as much as possible, and it is observed that certain post-merge point positions are conducive to greater simplification, intuitively by better preparing for future simplifications as more merge operators will be valid thereafter.

We will now specify the constraints (topology and classification) that will make it possible to determine whether a merge operator during simulation is applicable ("valid") or not.

For the preservation constraint of the topology, we can use three notions: A "link condition", described in [Ref21], for the D-manifold property. In 2D this condition ensures that any neighborhood of a point is homeomorphic (topologically equivalent) to a disk or a half-disk. On the topological plane this condition requires that each edge is adjacent to exactly one or two triangular faces, and that each vertex is adjacent to a single connected component. In 3D this condition ensures that any neighborhood of a point is homeomorphic (topologically equivalent) to a ball or half-ball.

The determination of a polyhedron kernel to verify that the embedding is valid [Ref27]. In 3D the nucleus of a polyhedron is the place of the points which see the whole edge of the polyhedron. In the general case the nucleus is either empty or a convex polyhedron. The polyhedron considered for the calculation of the nucleus is the polyhedron formed by the union of the tetrahedra adjacent to the two vertices of the edge. When the nucleus is non-empty, we can merge the edge by placing the target vertex in the nucleus: the triangulation remains valid, that is to say without overlapping or inverting tetrahedra. When the kernel is empty you can not merge the edge without creating an invalid triangulation. In our context we use the construction of the nucleus to delimit a zone of search of points which classify well (this is the point below).

The preservation of the classification of the points of the sample, which is done as before, and which will be reviewed below.

A merge operator being simulated will therefore be considered valid if:

The link condition is verified,

The valid embedding condition is verified, the classification of the points of the sample is preserved. The simplification process may be in accordance with Figure 17.

We consider one by one all the edges existing in different parts of the triangulation, starting with the edge of the simplistic tolerance, as will be seen.

For each edge, step 1701 first simulates the fusion of this edge, to determine if this fusion is possible ("valid"). As already indicated, we will first try a half-edge merge operator, then a general edge merge operator. The simulation is based on the current state of the triangulation. If the merge operator being simulated is valid, step 1703 stores the definition in a dynamic priority queue. The order of priority is defined by the aforementioned optimization criterion.

In step 1705, the first merge operator is applied in the order of the priority queue.

After that, step 1707 updates the priority queue: deletion of operators that are no longer required because they relate to an edge deleted by the merge operator that has just been executed.

Re-simulation of modified operators, which are put back in the queue, with their optimization criteria recalculated. An operator is "modified" when the merge operator that has just been executed has modified the parts of the triangulation that are adjacent to at least one of the two vertices of the edge on which it is carrying (or both vertices ).

Sub-steps 2072 to 2076 of step 207 of FIG. 4 show that the operation is carried out in three stages. At each time, half-edge fusions are first considered, followed by general edge fusions.

Here, a reciprocal triangulation (described in detail later) takes place in step 2074, within the simplification phase. It would be conceivable to perform this reciprocal triangulation just after doing the refinement. In other words, one can imagine simplifying all possible edges after doing all the refinement, including reciprocal triangulation. The proposed process is considered more parsimonious because it allows a lower complexity of the triangulation.

Thus, the Applicant currently believes that it is preferable to simplify the triangulation of the tolerance volume (the simplistic tolerance) before launching the reciprocal triangulation, then the rest of the simplification. The idea is to reduce the size of the triangulation before simplifying it because the number of operators to use depends on this size.

Also note the following point: in refinement, the proposed 3D triangulation is that of Delaunay; it is convenient because it is canonical for a set of points, and we know how to prove the termination of refinement. But it is "verbose": for it is at fixed points a triangulation the most isotropic possible- so to speak little parsimonious (in points) in places where the geometry is anisotropic. The idea is to reduce the size of the triangulation before simplifying it because the number of operators to use depends on this size. The simplification will now be illustrated with reference to Figures 8.0 to 8.11.

Figure 8.0 shows the mesh, in its state of figure 7.7, whose figure 7.8 showed that it was the simplistic tolerance Γ. We know that the isosurface Zset is contained in this volume of "simple tolerance" noted Γ. The simplification first concerns the edge of the "simplistic tolerance" (step 2072 of FIG. 4).

In Figure 8.0, two of the points are connected by an edge SF1, shown in bold, to indicate that they can be merged.

As already indicated, the merger between two adjacent vertices makes disappear the edge which joins them, one of these two vertices, and the faces or tetrahedra adjacent to the removed edge.

When we merge two vertices in 3D, we will also merge (collapse one on the other) two edges located between these two points and a third (which is not in the plane of the figure). Therefore, we must consider all the primitives (2D triangles or 3D tetrahedra) that are adjacent to the vertices of the removed edge. The points in the sample that are located in these primitives are called here "merge points".

It is then necessary that, in the vicinity of the edges to collapse, the error ε in all the points concerned by the fusion satisfies the aforementioned condition: ε <= (1- a) in other words that all the points concerned by the fusion are well classified (symbolized by squares) after the merge. This is called preserving the condition of classification of the points of the set S. As indicated, we determine in advance which fusion operators preserve the classification and the topology by "simulating" them, without apply them. Only the merge operators selected as suitable ("valid") after simulation are stored in the dynamic priority queue of step 1703.

In order to speed up the processing, half-edge collapse is done first, whose operators are much faster to simulate. Then, when no half-edge fusion is possible, we consider the more general operators of edge fusion. It can also be said that the Delaunay meshes concerned must be homogeneous as regards the good classification of the points. Note that at the level of simplification, it is not verified that the mesh continues to satisfy the Delaunay conditions. The result after this first fusion of half-edges and / or edges is illustrated in Figure 8.1.

We have redone the classification calculation, and the error ε, for those points of the set S that are concerned by the merger. The isosurface was also recalculated.

In Figure 8.2, two of the points are connected by an edge SF2, shown in bold, to indicate that they can be fused as well.

The result after this other merge is shown in Figure 8.3.

This is repeated with all the points that can be merged. When there are no more edges to merge (points that can be merged) at the edge of the oT simple tolerance, the result is shown in Figure 8.4. In summary, so far simplification has involved the following steps: hl. Delimit a volume called "simplistic tolerance", noted Γ. It consists of a mesh union of the Delaunay 3D triangulation, chosen to form a mesh approximation of the tolerance volume. This is illustrated in 2D in Figure 7.8, where the mixed line indicates the edge oT this simple tolerance. - h2. Perform edge fusions in oT, verifying the preservation of the S classification condition (Figures 8.1 to 8.4).

We continue as follows: h3. Perform a mutual triangulation between the isosurface (Zset) and the so-called "simplistic tolerance" volume, as modified in step h2 (step 2074 of FIG. 4). This mutual triangulation (sometimes called reciprocal triangulation) adds vertices with the value F (v) = 0 (where v denotes a vertex), edges and meshes in the 3D triangulation, on the isosurface.

By this mutual triangulation operation of step h3, we will somehow integrate the isosurface Z to the general triangulation of the volume of "simplistic tolerance", noted Γ.

In 2D the mutual triangulation between a 2D triangulation T and a polygonal line L (whose vertices coincide with the edges of the triangulation) consists to insert all the vertices of L in T, all the edges of L in T, then to conform the triangulation by adding edges to T to obtain only triangles.

The operation of the mutual triangulation is illustrated in Figures 15, on an object of the same pace as that of Figure 3: At the beginning (Figure 15.1), there is the existing triangulation (at the point where it is rendered) in hatched lines , and the Zset isosurface in bold lines.

Figure 15.2 illustrates the reciprocal triangulation of the Zset curve and the existing triangulation: new vertices (small black disks) appear on the Zset line, new edges, and new triangles. The Zset curve is now represented as a chain of edges of the triangulation.

Finally, in Figure 15.3, the triangles are classified according to the sign of the interpolated function (white = positive, gray = negative).

For the example of FIG. 8, the result is illustrated in FIG. 8.5. There are edges added to the hatched triangulation, and new vertices (small black disks). h4. Make fusions of edges on this isosurface Z, which takes its final configuration, after these simplifications. (Figures 8.6 to 8.8). h5. Continue to perform edge fusions on all possible edges in the "simplistic tolerance" volume, including checking the preservation of the S-point classification condition of sample points (step 2076 of Figure 4).

The process continues, as long as there are edges to merge, on the inner side of Z. It stops when no edge can be merged without violating the conditions for classifying points and topology. In the above, the simplification uses edge merge operators as triangulation modification operators. It is conceivable to use other triangulation modification operators, at least in part, for example: continuous operators (vertex displacement), combinatorial operators, which, in addition to the "half-edge" merger, already cited, include operators such as those called "4-4 flips", or "edge-removal", as shown in Figure 1 since http://graphics.berkelev.edu/papers/Klingner-ATM-2007-10/Klingner-ATM-2007- 10. pdf and mixed operators whose general edge fusion is an example.

It is also possible to use combinations of operators (for example general edge fusion, followed by a 4-4 flip, then an edge-removal).

The described embodiment avoids excessive requirements in terms of memory and computation complexity. Its final result is shown in Figure 8.10. Figure 8.11 is similar to Figure 8.10, but has been removed from the symbols illustrating the classification of points. In the final contour of Figure 8.11, the Zset isosurface fits exactly within the tolerance volume.

In the final, the Zset isosurface forms a meshed surface approximation of the tolerance volume. We can consider that Zset and oT are joined during the mutual triangulation. The mechanism just described (refinement + simplification) works well in the general case. It may have to be arranged in particular situations.

Some factors could compromise the refinement phase of the topology.

First, it is possible that the Zset isosurface crosses 5Ω between two samples, for example due to the presence of a flat tetrahedron ("sliver"). This is at least partly avoided by using the above-mentioned parameter in the classification. We have seen that a point s is considered as well classified if ε (s) <= (1 - a)

We then consider the heterogeneous tetrahedra, which are those whose vertices have different label values F (s). We also consider a magnitude or "height" h, which is the distance between the maximum dimension triangles, formed by the vertices of the tetrahedron that have the same label value.

In FIG. 9A, the three vertices B, C and D have the same label, and h is the minimum distance between the vertex A and the support plane of the BCD face.

In Figure 9B, the vertices A and B have the same label. Similarly, the vertices C and D have the same label, h is the minimum distance between the support line of AB and the CD support line. And we constrain these heterogeneous tetrahedra so that h> = 2 / (σα).

It is a way of making the interpolation function f satisfy the Lipschitz criterion, according to which the absolute value of the gradient of this function f is bounded.

A proper choice of made as well as the problems of flat tetrahedron ("sliver") are naturally solved.

As other disturbing factors of refinement, one can also encounter problems of orientation (perpendicular or "normal" direction) in complex areas.

They can be solved using the above. Preferably, and optionally, it will also be verified for each tetrahedron that an extension of the piecewise linear function outside the tetrahedron, also classifies points of the sample not only in the tetrahedron, but also outside. We will then say that the interpolation function f classifies the "local geometry".

The value to choose for a depends on the nature of the input data. The value a = 0.2 is suitable, at least as a starting value.

Reference is now made to Fig. 10, which depicts at the top of the picture crude representations of an elephant (from left to right: a cloud of points without noise, a soup of triangles without noise, a cloud of points with noise, and a soup of triangles with noise), and in the lower part the triangulation obtained according to the invention. For each example the raw data is converted to a tolerance volume by calculating a sub-level of the distance function to the data (points or triangles).

On the left (Figure 10A), the starting image is a cloud of points that have no noticeable effect on the final triangulated surface. Further to the right (Figure 10B), the starting data forms a soup of unorganized triangles (with intersections, holes) that also have no noticeable effect on the quality of the triangulated final surface. Further to the right (Figure 10 C), the starting data is a cloud of points with a high level of noise (uncertainty in their coordinates) which also has no noticeable effect on the quality of the triangulated final surface . Finally, quite right (Figure 10 D), the starting data form a very irregular and noisy triangles soup that also has no noticeable effect on the quality of the final triangulated surface. It is simply noted that the triangulation is less detailed in Figures 10C and 10D, where the starting data have a high noise level, and for which a larger tolerance volume is chosen.

These examples illustrate the robustness of the technique according to the invention. The invention has been developed using the following computer database:

INTEL Intel 2.4GHz 16-core machine with 128 gigabytes of RAM,

For the triangulation data structures, CGAL library, available from Geometry Factory, 20, Yves Emmanuel Baudoin Avenue, 06130 Grasse, France,

INTEL library "Threading Building Blocks library" for parallelization. It is observed that the "atomic" treatments on the tetrahedrons and the sample points are independent, and can therefore easily be conducted in parallel.

Fig. 11 shows the part of Figs. 1. Fig. 11.0 is the tolerance volume. Figures 11.1 to 11.4 illustrate the final grid obtained, with different processing conditions, with respect to the factor δ, which is the separation distance between the boundaries of the tolerance volume, expressed as a percentage. The following table shows the values of δ, the processing time with the equipment indicated above, the figure with the final mesh, and the number of vertices thereof.

It appears that the calculation time decreases when the tolerance increases. In general, the process is slow, which is a counterpart to its robustness.

The performances obtained by the invention are illustrated in FIG.

It will be observed that a strict comparison with the prior art is not possible since the constraints of the problem posed here are different.

The invention can however be compared to the following proposals: Lindstrom-Turk, which is the method described in [Refl9], which can be implemented using the CGAL library version 4.6, marketed by Geometry Factory already cited

MMGS, which is the method described in [Ref7], implemented by MMGS software version 2.0a, transmitted directly by Pascal Frey, second author of [Ref7], M MGS-A, which is a variant of the same M MGS software with the mesh anisotropy option,

Open Flipper, which is the process described in [Ref20], with software available for download from www.openfjipper.org/, version 2.1. - Simplification Envelopes, which is [Ref9], with software version 1.2, transmitted by email by the first author.

The Haussdorf distance measures the dissimilarity between two geometrical forms. The unidirectional distance Haussdorf (A, B) is defined by the maximum of the Euclidean distance between the form A and the form B, for every point of A, and vice versa for Haussdorf (B, A). The symmetrical variant is the maximum of Haussdorf (A, B) and Haussdorf (B, A).

The ordinate scale of Figure 12 corresponds to Hausdorff's distance H | o-> i, which is the unidirectional Hausdorff distance between the output and the input. The abscissa scale is the number of vertices at the output. This figure 12 shows that the invention is better than the other techniques by Haussdorf distances lower, with numbers of peaks in the final lower, in addition to the robustness already shown above.

Figures 12 and 13 have as input a gross surface mesh of the images of Figures 14 (about 14,000 vertices). Figure 13 shows the time of treatment. It illustrates the evolution of the mesh complexity as a function of processing and time for different values of the input tolerance δ, indicated in the legend rectangle. On the abscissa, we have the different stages of treatment, namely:

Refinement, - Simplification by half-edge mergers of the border dT of the simplistic tolerance,

Simplification by border fusions of the dT boundary of the simplest tolerance, mutual triangulation and zeroSet Z half-edge fusions, zeroSet Z edge fusions,

Simplification of all edges. The invention also makes it possible to treat surfaces having holes, for example when they are obtained by laser scanning which tangents concave zones. The holes can be deleted, as long as the tolerance volume includes them.

The invention can also be defined as a device. Figure 18 shows the elements of Figure 4, the steps are performed by operators. In numerical references, the most significant figure is 2 for steps, as opposed to 3 for operators.

Thus, the initialization step 201 of FIG. 4 is performed by the initialization operator 301 of FIG. 18. It starts from the tolerance volume, stored in 310. Inside this initialization, the step 2011 of Figure 4 is performed by the contour sampling operator 3011 of Figure 18.

Figure 19 illustrates a point classification operator, referenced 1900.

For a point to be classified, we start from the following data (at least partly, some data being deductible from the others): - the coordinates of this point, the label or label of this point, the mesh which surrounds this point, with the coordinates and labels of the vertices of this mesh.

A 1901 operator calculates the interpolation function f, as described above with reference to FIG. 5. It provides a label calculated for this point. The operator 1902 deduces whether the point is well classified or not, as already described.

It will be observed that this figure 19 does not reflect exceptional cases, such as the one where all the vertices of the mesh have the same label value.

This point classification operator, 1900, is called many times, in particular by the operators 3013 and 3032 of FIG. 18, as well as by the simulation operator 1701 of FIG. 17, insofar as the latter must check that the classification of points is preserved. Appendix 1 - Definitions

isomorphism

In mathematics, an isomorphism between two structured sets is a bijective application that preserves the structure and whose inverse also preserves the structure (For many structures in algebra, this second condition is automatically fulfilled, but this is not the case in topology for example where a bijection can be continuous, without its reciprocal being).

Hom (e) omorphism (in English: "homomorphism")

In topology, a homeomorphism is a continuous bijective application, of a topological space in another, whose reciprocal bijection is also continuous. In view of the above, it is an isomorphism between two topological spaces.

When we say that two topological spaces are homeomorphic, it means that they are "the same", seen differently. This does not always make sense. For example, we show that a cup (with handle) is homeomorphic to a torus, because there is a transformation that moves from one to the other: fill the body of the cup of matter, then deform it gradually, with its handle, so that the whole forms a torus. The reciprocal transformation restores the cup.

isotopy

The notion of isotopy is particularly important in node theory: two nodes are considered identical if they are isotopic, that is to say if one can deform one to obtain the other without the "rope" tears or does not penetrate.

This notion of isotopy should not be confused with isotopy in chemistry, nor with isotropy in physics of materials.

Tolerance volume (in English: "volume tolerance") This applies to a three-dimensional component. The tolerance volume for a component is the volume between its largest and the smallest allowable outer surface.

In practice, there may be several related components of the tolerance volume. There may also be several related components of its edge surfaces, even though the tolerance volume itself has only one related component. Annex 2 - References

[Refl]. AGARWAL, P. K. AND SURI, S. 1998. Surface approximation and geometry partitions. Journal of

Computing 27, 4, 1016-1035.

[Ref 2]. AMENTA, N., AND BERN, M. 1998. Surface reconstruction by voronoi filtering. Discreet and

Computational Geometry 22, 481-504.

[Ref3]. AMENTA, N., CHOI, S., DEY, T.K., AND LEE HA, N. 2000. A simple algorithm for homeomorphic surface reconstruction. In Proceedings of ACM Symposium on Computational Geometry, 213-222. [Ref4]. BISCHOFF, S., PAVIC, D. AND KOBBELT, L. 2005. Automatic restoration of polygon models. ACM Transactions on Graphics 24, 1332-1352.

[Ref5]. BOISSONNAT, J.-D., AND CAZALS, F. 2000. Smooth surface reconstruction via natural neighbor interpolation of distance functions. In Proceedings of ACM Symposium on Computational Geometry,

223-232.

[Ref6]. BOISSONNAT, J.-D., AND OUDOT, S. 2005. Provably good sampling and meshing of surfaces.

Graphical Models 67, 5, 405-451.

[Ref7]. BOROUCHA I, H., AND FREY, P. 2005. Simplification of a surface mesh using hausdorff envelope.

Computer Methods in Applied Mechanics and Engineering 194, 48-49, 4864-4884.

[Ref8]. CHAZAL, F., AND COHEN-STEINER, D. 2004. A condition for isotopic approximation. in

Proceedings of ACM Symposium on Solid Modeling and Applications, 93-99.

[Ref9]. COHEN, J., VARSHNEY, A., MANOCHA, D., TUR, G., WEBER, H., AGARWAL, P., BROO S, F., AND WRIGHT,

W. 1996. Simplification envelopes. In Proceedings of ACM Conference on Computer Graphics and Interactive Techniques, 119-128.

[REFLO]. COHEN, J., MANOCHA, D., AND OLANO, M. 2003. Successive mappings: An approach to polygonal mesh simplification with guaranteed error bounds. International Journal of Computational Geometry and Applications 13, 1, 61-96.

[Refll]. DEY, T. K., AND GOSWAMI, S. 2003. Tight cocoon: A watertight surface reconstructor. in

Proceedings of ACM Symposium on Solid Modeling and Applications, 127-134.

[Refl2]. DEY, T. K., AND SUN, J. 2005. An adaptive put surface for reconstruction with guarantees. In Proceedings of EUROGRAPHICS Symposium on Geometry Processing.

[Refl3]. DEY, T.K., Li, K., RAMOS, E.A., AND WENGER, R. 2009. Isotopic reconstruction of surfaces with boundaries. Computer Graphics Forum 28, 5, 1371-1382.

[Refl4]. DEY, T. K. 2006. Curve and Surface Reconstruction: Algorithm with Mathematical Analysis (Cambridge Monographs on Applied and Computational Mathematics). Cambridge University Press. [Refl5]. GUEZIEC, A. 1996. Surface simplification inside a tolerance volume. Tech. Rep. 20440. IBM Research Report RC 20440.

[Refl6]. HORNUNG, A., AND KOBBELT, L. 2006. Robust reconstruction of watertight models from non-uniformly sampled point clouds without normal information. In Proceedings of EUROGRAPHICS Symposium on Geometry Processing, 41-50.

[Refl7]. Ju, T. 2004. Robust repair of polygonal models. ACM Transactions on Graphics 23, 3, 888- 895.

[Refl8]. KALVIN, A.D., AND TAYLOR, R.H. 1996. Superfaces: Polygonal mesh simplification with

bounded error. IEEE Computer Graphics and Applications 16, 3.

[Refl9]. LINDSTROM, P., AND TURK, G. 1999. Evaluation of memoryless simplification. IEEE Transactions on Visualization and Computer Graphics 5, 2, 98-115.

[Ref20]. MOBIUS, J., AND KOBBELT, L. 2012. Openflipper: An open source geometry processing and rendering framework. In Proceedings of International Conference on Curves and Surfaces, Springer-

Verlag, 488-500.

[Ref21]. Tamal K. Dey, Anil N. Hirani, Bala Krishnamoorthy, Gavin Smith. Edge Contractions and Simplicial Homology. arXiv: 1304.0664.

[Ref22]. AGGARWAL, A., BOOTH, H., O'ROURKE, J., SURI, S., AND YAP, C. K. 1985. Finding minimal convex nested polygons. In Proceedings of ACM Symposium on Computational Geometry, 296-304.

[Ref23]. BOISSONNAT, J.-D., AND YVINEC, M. 1998. Algorithmic Geometry. Cambridge University Press. [Ref24]. CIAMPALINI, A., CIGNONI, P., MONTANI, C., AND SCOPIGNO, R. 1997. Multiresolution decimation based on global error. The Visual Computer 13.

[Ref25]. COXETER, H.S.M. 1969. Introduction to geometry (2nd ed.). John Wiley and Sons.

[Ref26]. GUMHOLD, S., BORODIN, P., AND KLEIN, R. 2003. Intersection free simplification. International

Journal of Shape Modeling 9, 2.

[Ref27]. Lee, D.T .; Preparata, F.P. (1979), "An Optimal Algorithm for Finding the Kernel of a

Polygon ", Journal of the ACM 26 (3): 415-421, doi: 10.1145 / 322139.322142.

[Ref28]. ZELINKA, S., AND GARLAND, M. 2002. Permission grids: Practical, error-bounded simplification.

ACM Transactions on Graphics 21, 2, 207-229.

Claims

claims

A method of processing geometric data, comprising the steps of:

A. receive a tolerance volume Ω relative to raw geometrical data of departure,

B. initiating (201) a canonical triangulation in the tolerance volume, from a set S of sample points taken at the limits of this tolerance volume, and points situated outside thereof,

C. refine (203) this canonical triangulation, to classify the points of the sample, which provides a dense mesh of tolerance volume, and

D. simplify (207) this dense mesh by triangulation modification operations, preserving the topology and the classification of the points of the sample.

A method according to claim 1, characterized in that step B comprises the sub-step (2011) of sampling the limits of the tolerance volume according to a selected sampling density σ, which provides the set S of sample points .

A method according to claim 2, characterized in that step B comprises marking (2013) each sample point of the set S with a tag value which represents the index of the edge connected component of that point, in the tolerance volume, and in that step C comprises the iterative generation of the canonical triangulation, by inserting one point at a time (2031), on at least a part of the set S of points, until that all the points concerned (2033) satisfy a classification condition (2032), based on the value of a function F itself dependent on their index value, and on its interpolation by an interpolation function f applied to each mesh of the 3D triangulation, and the construction of a piecewise linear isosurface Zset, using the points where the interpolation function f has the same selected intermediate value, in particular zero, this Zset isosurface being a closed surface who f form a surface approximation of the volume of tolerance.

4. Method according to one of the preceding claims, characterized in that the canonical triangulation is a Delaunay triangulation.

5. Method according to one of the preceding claims, characterized in that the initial raw geometric data are 3D data.

6. The method of processing geometric data according to one of the preceding claims, characterized in that step D. comprises the following operations:

Dl. determining (1701) whether triangulation modification operations are valid, i.e. preserve the topology of the mesh and preserve the condition of classifying the set S of sample points, and

D2. perform (1703-1709) valid triangulation change operations in a specified order.

Method according to claim 6, characterized in that the triangulation modification operations comprise edge fusions and / or half-edge fusions.

8. Method according to one of claims 6 and 7, characterized in that the order is determined according to an optimization criterion.

9. Method according to one of claims 6 and 7, characterized in that step D. is a volume called "simplistic tolerance" approaching the tolerance volume by a mesh of the canonical triangulation, the isosurface Zset being contained in this volume called "simplistic tolerance",

And in that the steps D1 and D2 are performed:

First (2072) on the edge of the volume called "simplistic tolerance",

Then on the fruit (2074) of a mutual triangulation between the isosurface (Zset) and the volume called "simplistic tolerance", this mutual triangulation adding vertices, edges and meshes in the triangulation, on the isosurface,

Finally (2076), with other edges contained in the volume called "simplistic tolerance", to cover all possible edges.

A device for processing geometric data, comprising:

A geometric data memory, for receiving data (310) representing a tolerance volume, An initialization member (301), arranged to perform a first canonical triangulation between the edge of the tolerance volume and points located outside thereof,

A refinement member (303), arranged to specify said first triangulation to provide a dense mesh of the tolerance volume, which preserves a classification condition of the points of the triangulation,

A simplification member (307), arranged to simplify said dense mesh, preserving the topology and the classification of its points, and

A driver (300) which successively actuates the initialization member (301), the refinement member (303), and the simplification member (307), for changing graphic data from the same tolerance volume .

11. Device according to claim 10, characterized in that it comprises: a point classification operator (1900), which receives the designation of a point to be treated, and determines an interpolation function on the basis of the mesh surrounding this point to be processed, and label values relative to the vertices of this mesh, to provide an interpolated label, and compare it to a current label of the point to be treated, which allows to determine a classification condition of the point to treat.

A computer program product comprising portions of program code for implementing the method according to one of claims 1 to 9, or the device according to one of claims 10 and 11, when the program is executed on a computer.